Study Of Feature Selection Method Based On Mutual Information

Pattern recognition is an important application in signal and processing fields. With the development of artificial intelligence in the 50th century, Pattern recognition developed speedily and applied widely. Methods and theories based on it had been applied in many scientific and technological fields, it's had advanced the development of artificial intelligence, expanded the usage of computer.Pattern recognition is to classify one thing to one class correctly. Pattern recognition methods based on statistical method is consist of data acquisition, preprocessing, feature abstraction and classify decision making.Feature selection plays an important role in pattern recognition. It will be a time consuming thing to classify directly using the dataset which is huge after feature abstracting. Feature selection is to find a good algorithm method to select a dataset which is available to classify with few data in permission time. There were many excellent feature selection algorithms such as optimal search algorithm, branch and bound feature selection algorithm, simulated annealing algorithm, Tubu searching algorithm, genetic algorithm Taguch experimental method and so on.Feature selection based on information theory is a new subject in recent years. The relationships not only between input features and output classes but also between the input features must be considered in feature selection based on information theory, the mutual information between an input feature and output classes indicates its role to classification, the mutual information between two input features indicates their redundant , feature selection based on mutual information is to find out some features with big mutual information to output classes and small mutual information to other input features. The proposed algorithms in the paper are applied to several classification problems and compared with two previous algorithms, the experiment results indicated that the new methods increase the feature selection performance, it can work well not only in linear classification problems but also in nonlinear classification ones...